blob: 37dab2898ad013bc6f9aec091935bc6a891e790d [file] [log] [blame]
Andy Grover1e23b3e2009-02-24 15:30:34 +00001/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090034#include <linux/slab.h>
Andy Grover1e23b3e2009-02-24 15:30:34 +000035#include <linux/pci.h>
36#include <linux/dma-mapping.h>
37#include <rdma/rdma_cm.h>
38
39#include "rds.h"
40#include "ib.h"
41
42static struct kmem_cache *rds_ib_incoming_slab;
43static struct kmem_cache *rds_ib_frag_slab;
44static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
45
Andy Grover1e23b3e2009-02-24 15:30:34 +000046void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
47{
48 struct rds_ib_recv_work *recv;
49 u32 i;
50
51 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
52 struct ib_sge *sge;
53
54 recv->r_ibinc = NULL;
55 recv->r_frag = NULL;
56
57 recv->r_wr.next = NULL;
58 recv->r_wr.wr_id = i;
59 recv->r_wr.sg_list = recv->r_sge;
60 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
61
Andy Grover919ced42010-01-13 16:32:24 -080062 sge = &recv->r_sge[0];
Andy Grover1e23b3e2009-02-24 15:30:34 +000063 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
64 sge->length = sizeof(struct rds_header);
65 sge->lkey = ic->i_mr->lkey;
Andy Grover919ced42010-01-13 16:32:24 -080066
67 sge = &recv->r_sge[1];
68 sge->addr = 0;
69 sge->length = RDS_FRAG_SIZE;
70 sge->lkey = ic->i_mr->lkey;
Andy Grover1e23b3e2009-02-24 15:30:34 +000071 }
72}
73
Chris Mason33244122010-05-26 22:05:37 -070074/*
75 * The entire 'from' list, including the from element itself, is put on
76 * to the tail of the 'to' list.
77 */
78static void list_splice_entire_tail(struct list_head *from,
79 struct list_head *to)
80{
81 struct list_head *from_last = from->prev;
82
83 list_splice_tail(from_last, to);
84 list_add_tail(from_last, to);
85}
86
87static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
88{
89 struct list_head *tmp;
90
91 tmp = xchg(&cache->xfer, NULL);
92 if (tmp) {
93 if (cache->ready)
94 list_splice_entire_tail(tmp, cache->ready);
95 else
96 cache->ready = tmp;
97 }
98}
99
100static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
101{
102 struct rds_ib_cache_head *head;
103 int cpu;
104
105 cache->percpu = alloc_percpu(struct rds_ib_cache_head);
106 if (!cache->percpu)
107 return -ENOMEM;
108
109 for_each_possible_cpu(cpu) {
110 head = per_cpu_ptr(cache->percpu, cpu);
111 head->first = NULL;
112 head->count = 0;
113 }
114 cache->xfer = NULL;
115 cache->ready = NULL;
116
117 return 0;
118}
119
120int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
121{
122 int ret;
123
124 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
125 if (!ret) {
126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
127 if (ret)
128 free_percpu(ic->i_cache_incs.percpu);
129 }
130
131 return ret;
132}
133
134static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
135 struct list_head *caller_list)
136{
137 struct rds_ib_cache_head *head;
138 int cpu;
139
140 for_each_possible_cpu(cpu) {
141 head = per_cpu_ptr(cache->percpu, cpu);
142 if (head->first) {
143 list_splice_entire_tail(head->first, caller_list);
144 head->first = NULL;
145 }
146 }
147
148 if (cache->ready) {
149 list_splice_entire_tail(cache->ready, caller_list);
150 cache->ready = NULL;
151 }
152}
153
154void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
155{
156 struct rds_ib_incoming *inc;
157 struct rds_ib_incoming *inc_tmp;
158 struct rds_page_frag *frag;
159 struct rds_page_frag *frag_tmp;
160 LIST_HEAD(list);
161
162 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
163 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
164 free_percpu(ic->i_cache_incs.percpu);
165
166 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
167 list_del(&inc->ii_cache_entry);
168 WARN_ON(!list_empty(&inc->ii_frags));
169 kmem_cache_free(rds_ib_incoming_slab, inc);
170 }
171
172 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
173 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
174 free_percpu(ic->i_cache_frags.percpu);
175
176 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
177 list_del(&frag->f_cache_entry);
178 WARN_ON(!list_empty(&frag->f_item));
179 kmem_cache_free(rds_ib_frag_slab, frag);
180 }
181}
182
183/* fwd decl */
184static void rds_ib_recv_cache_put(struct list_head *new_item,
185 struct rds_ib_refill_cache *cache);
186static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
187
188
189/* Recycle frag and attached recv buffer f_sg */
190static void rds_ib_frag_free(struct rds_ib_connection *ic,
191 struct rds_page_frag *frag)
192{
193 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
194
195 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
196}
197
198/* Recycle inc after freeing attached frags */
199void rds_ib_inc_free(struct rds_incoming *inc)
200{
201 struct rds_ib_incoming *ibinc;
202 struct rds_page_frag *frag;
203 struct rds_page_frag *pos;
204 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
205
206 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
207
208 /* Free attached frags */
209 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
210 list_del_init(&frag->f_item);
211 rds_ib_frag_free(ic, frag);
212 }
213 BUG_ON(!list_empty(&ibinc->ii_frags));
214
215 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
216 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
217}
218
Andy Grover1e23b3e2009-02-24 15:30:34 +0000219static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
220 struct rds_ib_recv_work *recv)
221{
222 if (recv->r_ibinc) {
223 rds_inc_put(&recv->r_ibinc->ii_inc);
224 recv->r_ibinc = NULL;
225 }
226 if (recv->r_frag) {
Andy Groverfc24f782010-05-25 11:20:09 -0700227 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
Chris Mason33244122010-05-26 22:05:37 -0700228 rds_ib_frag_free(ic, recv->r_frag);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000229 recv->r_frag = NULL;
230 }
231}
232
233void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
234{
235 u32 i;
236
237 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
238 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000239}
240
Chris Mason037f18a32010-05-26 21:45:06 -0700241static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
242 gfp_t slab_mask)
Chris Mason33244122010-05-26 22:05:37 -0700243{
244 struct rds_ib_incoming *ibinc;
245 struct list_head *cache_item;
246 int avail_allocs;
247
248 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
249 if (cache_item) {
250 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
251 } else {
252 avail_allocs = atomic_add_unless(&rds_ib_allocation,
253 1, rds_ib_sysctl_max_recv_allocation);
254 if (!avail_allocs) {
255 rds_ib_stats_inc(s_ib_rx_alloc_limit);
256 return NULL;
257 }
Chris Mason037f18a32010-05-26 21:45:06 -0700258 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
Chris Mason33244122010-05-26 22:05:37 -0700259 if (!ibinc) {
260 atomic_dec(&rds_ib_allocation);
261 return NULL;
262 }
263 }
264 INIT_LIST_HEAD(&ibinc->ii_frags);
265 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
266
267 return ibinc;
268}
269
Chris Mason037f18a32010-05-26 21:45:06 -0700270static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
271 gfp_t slab_mask, gfp_t page_mask)
Chris Mason33244122010-05-26 22:05:37 -0700272{
273 struct rds_page_frag *frag;
274 struct list_head *cache_item;
275 int ret;
276
277 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
278 if (cache_item) {
279 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
280 } else {
Chris Mason037f18a32010-05-26 21:45:06 -0700281 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
Chris Mason33244122010-05-26 22:05:37 -0700282 if (!frag)
283 return NULL;
284
285 ret = rds_page_remainder_alloc(&frag->f_sg,
Chris Mason037f18a32010-05-26 21:45:06 -0700286 RDS_FRAG_SIZE, page_mask);
Chris Mason33244122010-05-26 22:05:37 -0700287 if (ret) {
288 kmem_cache_free(rds_ib_frag_slab, frag);
289 return NULL;
290 }
291 }
292
293 INIT_LIST_HEAD(&frag->f_item);
294
295 return frag;
296}
297
Andy Grover1e23b3e2009-02-24 15:30:34 +0000298static int rds_ib_recv_refill_one(struct rds_connection *conn,
Chris Mason037f18a32010-05-26 21:45:06 -0700299 struct rds_ib_recv_work *recv, int prefill)
Andy Grover1e23b3e2009-02-24 15:30:34 +0000300{
301 struct rds_ib_connection *ic = conn->c_transport_data;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000302 struct ib_sge *sge;
303 int ret = -ENOMEM;
Chris Mason037f18a32010-05-26 21:45:06 -0700304 gfp_t slab_mask = GFP_NOWAIT;
305 gfp_t page_mask = GFP_NOWAIT;
306
307 if (prefill) {
308 slab_mask = GFP_KERNEL;
309 page_mask = GFP_HIGHUSER;
310 }
Andy Grover1e23b3e2009-02-24 15:30:34 +0000311
Chris Mason33244122010-05-26 22:05:37 -0700312 if (!ic->i_cache_incs.ready)
313 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
314 if (!ic->i_cache_frags.ready)
315 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
316
Andy Grover3427e852010-05-24 20:28:49 -0700317 /*
318 * ibinc was taken from recv if recv contained the start of a message.
319 * recvs that were continuations will still have this allocated.
320 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800321 if (!recv->r_ibinc) {
Chris Mason037f18a32010-05-26 21:45:06 -0700322 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
Chris Mason33244122010-05-26 22:05:37 -0700323 if (!recv->r_ibinc)
Andy Grover1e23b3e2009-02-24 15:30:34 +0000324 goto out;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000325 }
326
Andy Grover3427e852010-05-24 20:28:49 -0700327 WARN_ON(recv->r_frag); /* leak! */
Chris Mason037f18a32010-05-26 21:45:06 -0700328 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
Andy Grover3427e852010-05-24 20:28:49 -0700329 if (!recv->r_frag)
330 goto out;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000331
Andy Grover0b088e02010-05-24 20:12:41 -0700332 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
333 1, DMA_FROM_DEVICE);
334 WARN_ON(ret != 1);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000335
Andy Grover919ced42010-01-13 16:32:24 -0800336 sge = &recv->r_sge[0];
Andy Grover1e23b3e2009-02-24 15:30:34 +0000337 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
338 sge->length = sizeof(struct rds_header);
339
Andy Grover919ced42010-01-13 16:32:24 -0800340 sge = &recv->r_sge[1];
Andy Grover0b088e02010-05-24 20:12:41 -0700341 sge->addr = sg_dma_address(&recv->r_frag->f_sg);
342 sge->length = sg_dma_len(&recv->r_frag->f_sg);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000343
344 ret = 0;
345out:
346 return ret;
347}
348
349/*
350 * This tries to allocate and post unused work requests after making sure that
351 * they have all the allocations they need to queue received fragments into
Chris Mason33244122010-05-26 22:05:37 -0700352 * sockets.
Andy Grover1e23b3e2009-02-24 15:30:34 +0000353 *
354 * -1 is returned if posting fails due to temporary resource exhaustion.
355 */
Andy Groverb6fb0df2010-06-23 18:06:30 -0700356void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
Andy Grover1e23b3e2009-02-24 15:30:34 +0000357{
358 struct rds_ib_connection *ic = conn->c_transport_data;
359 struct rds_ib_recv_work *recv;
360 struct ib_recv_wr *failed_wr;
361 unsigned int posted = 0;
362 int ret = 0;
363 u32 pos;
364
Joe Perchesf64f9e72009-11-29 16:55:45 -0800365 while ((prefill || rds_conn_up(conn)) &&
366 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000367 if (pos >= ic->i_recv_ring.w_nr) {
368 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
369 pos);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000370 break;
371 }
372
373 recv = &ic->i_recvs[pos];
Chris Mason037f18a32010-05-26 21:45:06 -0700374 ret = rds_ib_recv_refill_one(conn, recv, prefill);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000375 if (ret) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000376 break;
377 }
378
379 /* XXX when can this fail? */
380 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
381 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
Andy Grover0b088e02010-05-24 20:12:41 -0700382 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
383 (long) sg_dma_address(&recv->r_frag->f_sg), ret);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000384 if (ret) {
385 rds_ib_conn_error(conn, "recv post on "
386 "%pI4 returned %d, disconnecting and "
387 "reconnecting\n", &conn->c_faddr,
388 ret);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000389 break;
390 }
391
392 posted++;
393 }
394
395 /* We're doing flow control - update the window. */
396 if (ic->i_flowctl && posted)
397 rds_ib_advertise_credits(conn, posted);
398
399 if (ret)
400 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000401}
402
Chris Mason33244122010-05-26 22:05:37 -0700403/*
404 * We want to recycle several types of recv allocations, like incs and frags.
405 * To use this, the *_free() function passes in the ptr to a list_head within
406 * the recyclee, as well as the cache to put it on.
407 *
408 * First, we put the memory on a percpu list. When this reaches a certain size,
409 * We move it to an intermediate non-percpu list in a lockless manner, with some
410 * xchg/compxchg wizardry.
411 *
412 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
413 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
414 * list_empty() will return true with one element is actually present.
415 */
416static void rds_ib_recv_cache_put(struct list_head *new_item,
417 struct rds_ib_refill_cache *cache)
Andy Grover1e23b3e2009-02-24 15:30:34 +0000418{
Chris Mason33244122010-05-26 22:05:37 -0700419 unsigned long flags;
420 struct rds_ib_cache_head *chp;
421 struct list_head *old;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000422
Chris Mason33244122010-05-26 22:05:37 -0700423 local_irq_save(flags);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000424
Chris Mason33244122010-05-26 22:05:37 -0700425 chp = per_cpu_ptr(cache->percpu, smp_processor_id());
426 if (!chp->first)
427 INIT_LIST_HEAD(new_item);
428 else /* put on front */
429 list_add_tail(new_item, chp->first);
430 chp->first = new_item;
431 chp->count++;
432
433 if (chp->count < RDS_IB_RECYCLE_BATCH_COUNT)
434 goto end;
435
436 /*
437 * Return our per-cpu first list to the cache's xfer by atomically
438 * grabbing the current xfer list, appending it to our per-cpu list,
439 * and then atomically returning that entire list back to the
440 * cache's xfer list as long as it's still empty.
441 */
442 do {
443 old = xchg(&cache->xfer, NULL);
444 if (old)
445 list_splice_entire_tail(old, chp->first);
446 old = cmpxchg(&cache->xfer, NULL, chp->first);
447 } while (old);
448
449 chp->first = NULL;
450 chp->count = 0;
451end:
452 local_irq_restore(flags);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000453}
454
Chris Mason33244122010-05-26 22:05:37 -0700455static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
Andy Grover1e23b3e2009-02-24 15:30:34 +0000456{
Chris Mason33244122010-05-26 22:05:37 -0700457 struct list_head *head = cache->ready;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000458
Chris Mason33244122010-05-26 22:05:37 -0700459 if (head) {
460 if (!list_empty(head)) {
461 cache->ready = head->next;
462 list_del_init(head);
463 } else
464 cache->ready = NULL;
465 }
Andy Grover1e23b3e2009-02-24 15:30:34 +0000466
Chris Mason33244122010-05-26 22:05:37 -0700467 return head;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000468}
469
470int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
471 size_t size)
472{
473 struct rds_ib_incoming *ibinc;
474 struct rds_page_frag *frag;
475 struct iovec *iov = first_iov;
476 unsigned long to_copy;
477 unsigned long frag_off = 0;
478 unsigned long iov_off = 0;
479 int copied = 0;
480 int ret;
481 u32 len;
482
483 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
484 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
485 len = be32_to_cpu(inc->i_hdr.h_len);
486
487 while (copied < size && copied < len) {
488 if (frag_off == RDS_FRAG_SIZE) {
489 frag = list_entry(frag->f_item.next,
490 struct rds_page_frag, f_item);
491 frag_off = 0;
492 }
493 while (iov_off == iov->iov_len) {
494 iov_off = 0;
495 iov++;
496 }
497
498 to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
499 to_copy = min_t(size_t, to_copy, size - copied);
500 to_copy = min_t(unsigned long, to_copy, len - copied);
501
502 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
Andy Grover0b088e02010-05-24 20:12:41 -0700503 "[%p, %u] + %lu\n",
Andy Grover1e23b3e2009-02-24 15:30:34 +0000504 to_copy, iov->iov_base, iov->iov_len, iov_off,
Andy Grover0b088e02010-05-24 20:12:41 -0700505 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000506
507 /* XXX needs + offset for multiple recvs per page */
Andy Grover0b088e02010-05-24 20:12:41 -0700508 ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
509 frag->f_sg.offset + frag_off,
Andy Grover1e23b3e2009-02-24 15:30:34 +0000510 iov->iov_base + iov_off,
511 to_copy);
512 if (ret) {
513 copied = ret;
514 break;
515 }
516
517 iov_off += to_copy;
518 frag_off += to_copy;
519 copied += to_copy;
520 }
521
522 return copied;
523}
524
525/* ic starts out kzalloc()ed */
526void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
527{
528 struct ib_send_wr *wr = &ic->i_ack_wr;
529 struct ib_sge *sge = &ic->i_ack_sge;
530
531 sge->addr = ic->i_ack_dma;
532 sge->length = sizeof(struct rds_header);
533 sge->lkey = ic->i_mr->lkey;
534
535 wr->sg_list = sge;
536 wr->num_sge = 1;
537 wr->opcode = IB_WR_SEND;
538 wr->wr_id = RDS_IB_ACK_WR_ID;
539 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
540}
541
542/*
543 * You'd think that with reliable IB connections you wouldn't need to ack
544 * messages that have been received. The problem is that IB hardware generates
545 * an ack message before it has DMAed the message into memory. This creates a
546 * potential message loss if the HCA is disabled for any reason between when it
547 * sends the ack and before the message is DMAed and processed. This is only a
548 * potential issue if another HCA is available for fail-over.
549 *
550 * When the remote host receives our ack they'll free the sent message from
551 * their send queue. To decrease the latency of this we always send an ack
552 * immediately after we've received messages.
553 *
554 * For simplicity, we only have one ack in flight at a time. This puts
555 * pressure on senders to have deep enough send queues to absorb the latency of
556 * a single ack frame being in flight. This might not be good enough.
557 *
558 * This is implemented by have a long-lived send_wr and sge which point to a
559 * statically allocated ack frame. This ack wr does not fall under the ring
560 * accounting that the tx and rx wrs do. The QP attribute specifically makes
561 * room for it beyond the ring size. Send completion notices its special
562 * wr_id and avoids working with the ring in that case.
563 */
Andy Grover8cbd9602009-04-01 08:20:20 +0000564#ifndef KERNEL_HAS_ATOMIC64
Andy Grover1e23b3e2009-02-24 15:30:34 +0000565static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
566 int ack_required)
567{
Andy Grover8cbd9602009-04-01 08:20:20 +0000568 unsigned long flags;
569
570 spin_lock_irqsave(&ic->i_ack_lock, flags);
571 ic->i_ack_next = seq;
572 if (ack_required)
573 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
574 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
575}
576
577static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
578{
579 unsigned long flags;
580 u64 seq;
581
582 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
583
584 spin_lock_irqsave(&ic->i_ack_lock, flags);
585 seq = ic->i_ack_next;
586 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
587
588 return seq;
589}
590#else
591static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
592 int ack_required)
593{
594 atomic64_set(&ic->i_ack_next, seq);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000595 if (ack_required) {
596 smp_mb__before_clear_bit();
597 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
598 }
599}
600
601static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
602{
603 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
604 smp_mb__after_clear_bit();
605
Andy Grover8cbd9602009-04-01 08:20:20 +0000606 return atomic64_read(&ic->i_ack_next);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000607}
Andy Grover8cbd9602009-04-01 08:20:20 +0000608#endif
609
Andy Grover1e23b3e2009-02-24 15:30:34 +0000610
611static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
612{
613 struct rds_header *hdr = ic->i_ack;
614 struct ib_send_wr *failed_wr;
615 u64 seq;
616 int ret;
617
618 seq = rds_ib_get_ack(ic);
619
620 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
621 rds_message_populate_header(hdr, 0, 0, 0);
622 hdr->h_ack = cpu_to_be64(seq);
623 hdr->h_credit = adv_credits;
624 rds_message_make_checksum(hdr);
625 ic->i_ack_queued = jiffies;
626
627 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
628 if (unlikely(ret)) {
629 /* Failed to send. Release the WR, and
630 * force another ACK.
631 */
632 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
633 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
634
635 rds_ib_stats_inc(s_ib_ack_send_failure);
Andy Grover735f61e2010-03-11 13:49:55 +0000636
637 rds_ib_conn_error(ic->conn, "sending ack failed\n");
Andy Grover1e23b3e2009-02-24 15:30:34 +0000638 } else
639 rds_ib_stats_inc(s_ib_ack_sent);
640}
641
642/*
643 * There are 3 ways of getting acknowledgements to the peer:
644 * 1. We call rds_ib_attempt_ack from the recv completion handler
645 * to send an ACK-only frame.
646 * However, there can be only one such frame in the send queue
647 * at any time, so we may have to postpone it.
648 * 2. When another (data) packet is transmitted while there's
649 * an ACK in the queue, we piggyback the ACK sequence number
650 * on the data packet.
651 * 3. If the ACK WR is done sending, we get called from the
652 * send queue completion handler, and check whether there's
653 * another ACK pending (postponed because the WR was on the
654 * queue). If so, we transmit it.
655 *
656 * We maintain 2 variables:
657 * - i_ack_flags, which keeps track of whether the ACK WR
658 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
659 * - i_ack_next, which is the last sequence number we received
660 *
661 * Potentially, send queue and receive queue handlers can run concurrently.
Andy Grover8cbd9602009-04-01 08:20:20 +0000662 * It would be nice to not have to use a spinlock to synchronize things,
663 * but the one problem that rules this out is that 64bit updates are
664 * not atomic on all platforms. Things would be a lot simpler if
665 * we had atomic64 or maybe cmpxchg64 everywhere.
Andy Grover1e23b3e2009-02-24 15:30:34 +0000666 *
667 * Reconnecting complicates this picture just slightly. When we
668 * reconnect, we may be seeing duplicate packets. The peer
669 * is retransmitting them, because it hasn't seen an ACK for
670 * them. It is important that we ACK these.
671 *
672 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
673 * this flag set *MUST* be acknowledged immediately.
674 */
675
676/*
677 * When we get here, we're called from the recv queue handler.
678 * Check whether we ought to transmit an ACK.
679 */
680void rds_ib_attempt_ack(struct rds_ib_connection *ic)
681{
682 unsigned int adv_credits;
683
684 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
685 return;
686
687 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
688 rds_ib_stats_inc(s_ib_ack_send_delayed);
689 return;
690 }
691
692 /* Can we get a send credit? */
Steve Wise7b70d032009-04-09 14:09:39 +0000693 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000694 rds_ib_stats_inc(s_ib_tx_throttle);
695 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
696 return;
697 }
698
699 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
700 rds_ib_send_ack(ic, adv_credits);
701}
702
703/*
704 * We get here from the send completion handler, when the
705 * adapter tells us the ACK frame was sent.
706 */
707void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
708{
709 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
710 rds_ib_attempt_ack(ic);
711}
712
713/*
714 * This is called by the regular xmit code when it wants to piggyback
715 * an ACK on an outgoing frame.
716 */
717u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
718{
719 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
720 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
721 return rds_ib_get_ack(ic);
722}
723
724/*
725 * It's kind of lame that we're copying from the posted receive pages into
726 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
727 * them. But receiving new congestion bitmaps should be a *rare* event, so
728 * hopefully we won't need to invest that complexity in making it more
729 * efficient. By copying we can share a simpler core with TCP which has to
730 * copy.
731 */
732static void rds_ib_cong_recv(struct rds_connection *conn,
733 struct rds_ib_incoming *ibinc)
734{
735 struct rds_cong_map *map;
736 unsigned int map_off;
737 unsigned int map_page;
738 struct rds_page_frag *frag;
739 unsigned long frag_off;
740 unsigned long to_copy;
741 unsigned long copied;
742 uint64_t uncongested = 0;
743 void *addr;
744
745 /* catch completely corrupt packets */
746 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
747 return;
748
749 map = conn->c_fcong;
750 map_page = 0;
751 map_off = 0;
752
753 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
754 frag_off = 0;
755
756 copied = 0;
757
758 while (copied < RDS_CONG_MAP_BYTES) {
759 uint64_t *src, *dst;
760 unsigned int k;
761
762 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
763 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
764
Andy Grover0b088e02010-05-24 20:12:41 -0700765 addr = kmap_atomic(sg_page(&frag->f_sg), KM_SOFTIRQ0);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000766
767 src = addr + frag_off;
768 dst = (void *)map->m_page_addrs[map_page] + map_off;
769 for (k = 0; k < to_copy; k += 8) {
770 /* Record ports that became uncongested, ie
771 * bits that changed from 0 to 1. */
772 uncongested |= ~(*src) & *dst;
773 *dst++ = *src++;
774 }
775 kunmap_atomic(addr, KM_SOFTIRQ0);
776
777 copied += to_copy;
778
779 map_off += to_copy;
780 if (map_off == PAGE_SIZE) {
781 map_off = 0;
782 map_page++;
783 }
784
785 frag_off += to_copy;
786 if (frag_off == RDS_FRAG_SIZE) {
787 frag = list_entry(frag->f_item.next,
788 struct rds_page_frag, f_item);
789 frag_off = 0;
790 }
791 }
792
793 /* the congestion map is in little endian order */
794 uncongested = le64_to_cpu(uncongested);
795
796 rds_cong_map_updated(map, uncongested);
797}
798
799/*
800 * Rings are posted with all the allocations they'll need to queue the
801 * incoming message to the receiving socket so this can't fail.
802 * All fragments start with a header, so we can make sure we're not receiving
803 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
804 */
805struct rds_ib_ack_state {
806 u64 ack_next;
807 u64 ack_recv;
808 unsigned int ack_required:1;
809 unsigned int ack_next_valid:1;
810 unsigned int ack_recv_valid:1;
811};
812
813static void rds_ib_process_recv(struct rds_connection *conn,
Andy Grover597ddd52009-07-17 13:13:27 +0000814 struct rds_ib_recv_work *recv, u32 data_len,
Andy Grover1e23b3e2009-02-24 15:30:34 +0000815 struct rds_ib_ack_state *state)
816{
817 struct rds_ib_connection *ic = conn->c_transport_data;
818 struct rds_ib_incoming *ibinc = ic->i_ibinc;
819 struct rds_header *ihdr, *hdr;
820
821 /* XXX shut down the connection if port 0,0 are seen? */
822
823 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
Andy Grover597ddd52009-07-17 13:13:27 +0000824 data_len);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000825
Andy Grover597ddd52009-07-17 13:13:27 +0000826 if (data_len < sizeof(struct rds_header)) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000827 rds_ib_conn_error(conn, "incoming message "
828 "from %pI4 didn't inclue a "
829 "header, disconnecting and "
830 "reconnecting\n",
831 &conn->c_faddr);
832 return;
833 }
Andy Grover597ddd52009-07-17 13:13:27 +0000834 data_len -= sizeof(struct rds_header);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000835
Andy Groverf147dd92010-01-13 15:50:09 -0800836 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
Andy Grover1e23b3e2009-02-24 15:30:34 +0000837
838 /* Validate the checksum. */
839 if (!rds_message_verify_checksum(ihdr)) {
840 rds_ib_conn_error(conn, "incoming message "
841 "from %pI4 has corrupted header - "
842 "forcing a reconnect\n",
843 &conn->c_faddr);
844 rds_stats_inc(s_recv_drop_bad_checksum);
845 return;
846 }
847
848 /* Process the ACK sequence which comes with every packet */
849 state->ack_recv = be64_to_cpu(ihdr->h_ack);
850 state->ack_recv_valid = 1;
851
852 /* Process the credits update if there was one */
853 if (ihdr->h_credit)
854 rds_ib_send_add_credits(conn, ihdr->h_credit);
855
Andy Grover597ddd52009-07-17 13:13:27 +0000856 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000857 /* This is an ACK-only packet. The fact that it gets
858 * special treatment here is that historically, ACKs
859 * were rather special beasts.
860 */
861 rds_ib_stats_inc(s_ib_ack_received);
862
863 /*
864 * Usually the frags make their way on to incs and are then freed as
865 * the inc is freed. We don't go that route, so we have to drop the
866 * page ref ourselves. We can't just leave the page on the recv
867 * because that confuses the dma mapping of pages and each recv's use
Andy Grover0b088e02010-05-24 20:12:41 -0700868 * of a partial page.
Andy Grover1e23b3e2009-02-24 15:30:34 +0000869 *
870 * FIXME: Fold this into the code path below.
871 */
Chris Mason33244122010-05-26 22:05:37 -0700872 rds_ib_frag_free(ic, recv->r_frag);
Andy Grover0b088e02010-05-24 20:12:41 -0700873 recv->r_frag = NULL;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000874 return;
875 }
876
877 /*
878 * If we don't already have an inc on the connection then this
879 * fragment has a header and starts a message.. copy its header
880 * into the inc and save the inc so we can hang upcoming fragments
881 * off its list.
882 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800883 if (!ibinc) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000884 ibinc = recv->r_ibinc;
885 recv->r_ibinc = NULL;
886 ic->i_ibinc = ibinc;
887
888 hdr = &ibinc->ii_inc.i_hdr;
889 memcpy(hdr, ihdr, sizeof(*hdr));
890 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
891
892 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
893 ic->i_recv_data_rem, hdr->h_flags);
894 } else {
895 hdr = &ibinc->ii_inc.i_hdr;
896 /* We can't just use memcmp here; fragments of a
897 * single message may carry different ACKs */
Joe Perchesf64f9e72009-11-29 16:55:45 -0800898 if (hdr->h_sequence != ihdr->h_sequence ||
899 hdr->h_len != ihdr->h_len ||
900 hdr->h_sport != ihdr->h_sport ||
901 hdr->h_dport != ihdr->h_dport) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000902 rds_ib_conn_error(conn,
903 "fragment header mismatch; forcing reconnect\n");
904 return;
905 }
906 }
907
908 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
909 recv->r_frag = NULL;
910
911 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
912 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
913 else {
914 ic->i_recv_data_rem = 0;
915 ic->i_ibinc = NULL;
916
917 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
918 rds_ib_cong_recv(conn, ibinc);
919 else {
920 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
921 &ibinc->ii_inc, GFP_ATOMIC,
922 KM_SOFTIRQ0);
923 state->ack_next = be64_to_cpu(hdr->h_sequence);
924 state->ack_next_valid = 1;
925 }
926
927 /* Evaluate the ACK_REQUIRED flag *after* we received
928 * the complete frame, and after bumping the next_rx
929 * sequence. */
930 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
931 rds_stats_inc(s_recv_ack_required);
932 state->ack_required = 1;
933 }
934
935 rds_inc_put(&ibinc->ii_inc);
936 }
937}
938
939/*
940 * Plucking the oldest entry from the ring can be done concurrently with
941 * the thread refilling the ring. Each ring operation is protected by
942 * spinlocks and the transient state of refilling doesn't change the
943 * recording of which entry is oldest.
944 *
945 * This relies on IB only calling one cq comp_handler for each cq so that
946 * there will only be one caller of rds_recv_incoming() per RDS connection.
947 */
948void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
949{
950 struct rds_connection *conn = context;
951 struct rds_ib_connection *ic = conn->c_transport_data;
Andy Grover1e23b3e2009-02-24 15:30:34 +0000952
953 rdsdebug("conn %p cq %p\n", conn, cq);
954
955 rds_ib_stats_inc(s_ib_rx_cq_call);
956
Andy Groverd521b632009-10-30 08:51:57 +0000957 tasklet_schedule(&ic->i_recv_tasklet);
958}
Andy Grover1e23b3e2009-02-24 15:30:34 +0000959
Andy Groverd521b632009-10-30 08:51:57 +0000960static inline void rds_poll_cq(struct rds_ib_connection *ic,
961 struct rds_ib_ack_state *state)
962{
963 struct rds_connection *conn = ic->conn;
964 struct ib_wc wc;
965 struct rds_ib_recv_work *recv;
966
967 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000968 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
969 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
970 be32_to_cpu(wc.ex.imm_data));
971 rds_ib_stats_inc(s_ib_rx_cq_event);
972
973 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
974
Andy Groverfc24f782010-05-25 11:20:09 -0700975 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000976
977 /*
978 * Also process recvs in connecting state because it is possible
979 * to get a recv completion _before_ the rdmacm ESTABLISHED
980 * event is processed.
981 */
Zach Brownd455ab62010-07-06 15:04:34 -0700982 if (wc.status == IB_WC_SUCCESS) {
983 rds_ib_process_recv(conn, recv, wc.byte_len, state);
984 } else {
Andy Grover1e23b3e2009-02-24 15:30:34 +0000985 /* We expect errors as the qp is drained during shutdown */
Zach Brownd455ab62010-07-06 15:04:34 -0700986 if (rds_conn_up(conn) || rds_conn_connecting(conn))
Andy Grover1e23b3e2009-02-24 15:30:34 +0000987 rds_ib_conn_error(conn, "recv completion on "
Zach Brownd455ab62010-07-06 15:04:34 -0700988 "%pI4 had status %u, disconnecting and "
989 "reconnecting\n", &conn->c_faddr,
990 wc.status);
Andy Grover1e23b3e2009-02-24 15:30:34 +0000991 }
992
Zach Brownd455ab62010-07-06 15:04:34 -0700993 /*
994 * It's very important that we only free this ring entry if we've truly
995 * freed the resources allocated to the entry. The refilling path can
996 * leak if we don't.
997 */
Andy Grover1e23b3e2009-02-24 15:30:34 +0000998 rds_ib_ring_free(&ic->i_recv_ring, 1);
999 }
Andy Groverd521b632009-10-30 08:51:57 +00001000}
1001
1002void rds_ib_recv_tasklet_fn(unsigned long data)
1003{
1004 struct rds_ib_connection *ic = (struct rds_ib_connection *) data;
1005 struct rds_connection *conn = ic->conn;
1006 struct rds_ib_ack_state state = { 0, };
1007
1008 rds_poll_cq(ic, &state);
1009 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
1010 rds_poll_cq(ic, &state);
Andy Grover1e23b3e2009-02-24 15:30:34 +00001011
1012 if (state.ack_next_valid)
1013 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
1014 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
1015 rds_send_drop_acked(conn, state.ack_recv, NULL);
1016 ic->i_ack_recv = state.ack_recv;
1017 }
1018 if (rds_conn_up(conn))
1019 rds_ib_attempt_ack(ic);
1020
1021 /* If we ever end up with a really empty receive ring, we're
1022 * in deep trouble, as the sender will definitely see RNR
1023 * timeouts. */
1024 if (rds_ib_ring_empty(&ic->i_recv_ring))
1025 rds_ib_stats_inc(s_ib_rx_ring_empty);
1026
Andy Grover1e23b3e2009-02-24 15:30:34 +00001027 if (rds_ib_ring_low(&ic->i_recv_ring))
Andy Groverf17a1a52010-03-18 17:19:52 -07001028 rds_ib_recv_refill(conn, 0);
Andy Grover1e23b3e2009-02-24 15:30:34 +00001029}
1030
1031int rds_ib_recv(struct rds_connection *conn)
1032{
1033 struct rds_ib_connection *ic = conn->c_transport_data;
1034 int ret = 0;
1035
1036 rdsdebug("conn %p\n", conn);
Andy Grover1e23b3e2009-02-24 15:30:34 +00001037 if (rds_conn_up(conn))
1038 rds_ib_attempt_ack(ic);
1039
1040 return ret;
1041}
1042
1043int __init rds_ib_recv_init(void)
1044{
1045 struct sysinfo si;
1046 int ret = -ENOMEM;
1047
1048 /* Default to 30% of all available RAM for recv memory */
1049 si_meminfo(&si);
1050 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1051
1052 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1053 sizeof(struct rds_ib_incoming),
Andy Groverc20f5b92010-07-07 16:46:26 -07001054 0, SLAB_HWCACHE_ALIGN, NULL);
Andy Grover8690bfa2010-01-12 11:56:44 -08001055 if (!rds_ib_incoming_slab)
Andy Grover1e23b3e2009-02-24 15:30:34 +00001056 goto out;
1057
1058 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1059 sizeof(struct rds_page_frag),
Andy Groverc20f5b92010-07-07 16:46:26 -07001060 0, SLAB_HWCACHE_ALIGN, NULL);
Andy Grover8690bfa2010-01-12 11:56:44 -08001061 if (!rds_ib_frag_slab)
Andy Grover1e23b3e2009-02-24 15:30:34 +00001062 kmem_cache_destroy(rds_ib_incoming_slab);
1063 else
1064 ret = 0;
1065out:
1066 return ret;
1067}
1068
1069void rds_ib_recv_exit(void)
1070{
1071 kmem_cache_destroy(rds_ib_incoming_slab);
1072 kmem_cache_destroy(rds_ib_frag_slab);
1073}